Microsoft Researchers Expose Critical AI Vulnerability That Bypasses Safety Measures With Single Malicious Prompt

Microsoft researchers discover a devastating AI vulnerability called 'GRP-Obliteration' that completely bypasses safety measures across 15 major language models from OpenAI, Google, and Meta using just a single malicious prompt, forcing AI systems to prioritize compliance over safety and generate harmful content.

• Microsoft researchers discover that AI safety measures can be completely undone with just a single malicious prompt using a technique they call 'GRP-Obliteration'
• The method exploits Group Relative Policy Optimization training by changing what the judge model rewards, causing AI systems to learn compliance over safety and produce harmful content
• Microsoft successfully demonstrates this vulnerability across 15 major language models from companies including OpenAI, Google, Meta, and others, as well as image generation models